Abstracts

RETURN TO ABSTRACT LISTING


E2184. Utilization of a Novel Deep Learning Model in the Differentiation of Benign and Malignant Sonographic Breast Masses
Authors
  1. Aileen Chang; Santa Clara Valley Medical Center
  2. Pradnya Patel; Santa Clara University
  3. Christopher Nguyen; Santa Clara Valley Medical Center
  4. Yuling Yan; Santa Clara University
  5. Ran Pang; Santa Clara Valley Medical Center
  6. Mahesh Patel; Santa Clara Valley Medical Center
  7. Young Kang; Santa Clara Valley Medical Center
Objective:
To develop a deep learning model using a convolutional neural network to differentiate between benign and malignant sonographic breast masses.

Materials and Methods:
A novel deep learning model using a convolutional neural network (CNN) was constructed, consisting of 20 total layers (including 5 layers of convolution, 5 maxpool layers with ReLu as the activation function for feature extraction, 1 flatten layer, 4 batch normalization with fully connected layers, and 1 output layer). The CNN model was trained and validated on an open source breast ultrasound dataset by Al-Dhabyani et al. containing 780 images categorized into three classes: 133 normal, 487 benign, and 210 malignant images. 90% of the open source dataset was used for training the model; the remaining 10% was used to test the model, which achieved an accuracy of 94.2%. The deep learning model was then applied to 300 retrospectively gathered sonographic images of breast masses that were previously biopsied and received a pathologic diagnosis at our institution. The CNN was tested on this dataset, which classified the results as either “benign” or “malignant”, and its diagnostic performance was calculated based on the test set’s pathologic diagnosis.

Results:
Of the 300 masses in the test dataset, 194 were benign (64.7%) and 106 were malignant (35.3%). The CNN model had a sensitivity of 88.6%, specificity of 93.3%, and accuracy of 91.7%.

Conclusion:
Our deep learning model demonstrated high accuracy on our test dataset. The high diagnostic performance of the CNN model on a small dataset suggests its ability to utilize a smaller database for training. Furthermore, preliminary results suggest that our deep learning model may play a helpful role in differentiating between benign and malignant sonographic breast masses.